I would recommend checking Francois Chollet's tutorial for Autoencoders.
I just used his Convolutional Autoencoder example by replacing the target data.
Therefore, this network is not compressing data but outputting the MNIST handwritten digits with a different style. That's a simple image to image translation.
I draw these digits with an image manipulation program:
After 50 epoch:
Some of the challenging ones:
See the project on